It’s the monthly round of one-to-ones, and the QA sheet has five scored calls for each agent. Agent A passed the closing check, telling the customer what happens next, on all five calls. Agent B missed it once. B leaves with a coaching plan, and A leaves with a thank-you.
Over the whole month, A skipped the closing check on 3 calls in 10, and B on 1 in 20. Nobody in the room could have known. Each verdict rested on five calls, and five calls give an agent like A a clean sheet about one month in six.
No standard says how many calls QA should review per agent, and the most common sample is three to five a month. That is enough to find something to coach, and far too few to rank or compare agents. With five calls, an agent’s average can land 0.7 points either side of their true level. Telling two agents half a point apart takes about 40 scored calls each, and catching a failure that happens on 1 call in 100 takes every call. The right number depends on what you want the sample to answer.
Three to five calls is a habit with nothing behind it
Call Centre Helper says there are “no industry guidelines” for how many calls to monitor, and that the most common number evaluated per agent each month is 3 to 5. It puts the number down to how many QA staff a contact centre has budgeted for. What five calls can prove doesn’t come into it.
At five calls, each call is a fifth of an agent’s month. Say an agent scores 3.6 on four calls and 1.6 on one with a customer who wanted to argue. Their month averages 3.2, and they drop from On track to Coaching opportunity because of one conversation.
Five calls can’t separate a 3.2 agent from a 3.8 agent
An agent’s scores vary from call to call, because customers do. As an illustration, take a 1–5 scorecard and an agent whose call scores have a standard deviation of 0.8. You can check your own with one STDEV formula over last month’s scores for a single agent.
Average five of that agent’s calls, picked at random, and 95 times in 100 the result lands within 0.7 of their true level. An agent whose true level is 3.2 can show anywhere from 2.5 to 3.9, and one at 3.8 anywhere from 3.1 to 4.5. In about one month in eight, the 3.2 agent outscores the 3.8 agent.
Team figures are a different matter. Thirty agents with five calls each give the team average 150 calls, and that lands within about ±0.13 of the true level. The trouble is the per-agent number, the one that decides who gets coached.
More calls narrow the range, but slowly, because precision grows with the square root of the sample:
| Calls per agent per month | One agent’s average lands within | Gap between two agents you’d spot 4 times in 5 |
|---|---|---|
| 5 | ±0.7 | 1.4 |
| 10 | ±0.5 | 1.0 |
| 20 | ±0.35 | 0.7 |
| 40 | ±0.25 | 0.5 |
| 100 | ±0.16 | 0.3 |
Assumes a standard deviation of 0.8, calls picked at random, and 95% confidence. To work out your own number, take the precision you need and use calls ≈ (2 × standard deviation ÷ precision)². At 0.8 and ±0.5, that’s 10 calls. At ±0.25, it’s 41.
These are the best case. Real samples are rarely random: reviewers pick long calls, complaints and calls whose recordings play. And when different team leads score each agent’s five calls, their disagreement widens the spread further. Four team leads can score one call 2, 3, 3 and 5.
A clean sheet on a pass/fail check means less than it looks
Pass/fail checks, such as verifying identity or reading a required disclosure, behave worse, because one miss is the whole result.
Go back to the two agents. A passes the closing check on 7 calls in 10. The chance that five random calls all pass is 0.7⁵, or 17%: a clean sheet about one month in six, or two clean months in an average year. B passes on 19 calls in 20. The chance that five calls include at least one miss is 1 − 0.95⁵, or 23%: B gets flagged in nearly three months a year.
Over a year of five-call reviews, the agent who misses the check six times as often can look just as reliable on paper, and in some months more reliable.
A sample can’t find the failure that happens once in 100 calls
Some failures are rare and expensive: a missed sign that a customer is vulnerable, a wrong fee quoted, a complaint never logged. Say one happens on 1 call in 100 for a given agent. A random sample of five calls includes it about one month in 20. Ten calls catch it 10% of the time, 20 calls 18%, and 50 calls 39%.
No affordable manual sample will find it. Sampling describes an agent’s typical call well and their worst call badly, and the worst call is the one your compliance team, or your board, will ask about. That’s why we’ve written about evidencing good outcomes on every call.
Five calls are plenty to coach from
Five calls do one job well. A coaching conversation needs two or three calls the agent and team lead can listen to together: one where the skill slipped, and one where it went right. Picked on purpose, five calls give a coach more than enough to work with.
The trouble starts when the same five calls also produce the score that goes into the performance review, the team league table and the bonus. Choosing calls to teach from and sampling calls to measure are different jobs, and a sample chosen for one does the other badly.
Pick the number from the question
| What you want to know | Calls per agent per month | Why |
|---|---|---|
| What should we coach this month? | 3 to 5, picked on purpose | You need examples to listen to, and an average adds nothing |
| Where does this agent stand, give or take half a point? | About 10, at random | ±0.5 at a standard deviation of 0.8 |
| Is agent A really better than agent B? | About 40 each, at random | Spots a half-point gap 4 times in 5 |
| Did coaching move this agent’s score by half a point? | About 40 before and 40 after, plus a comparison group | Weak agents’ scores rise even when coaching does nothing |
| Does this ever happen? | Every call | A 1-in-100 failure slips past any sample you can afford |
The fourth row needs the comparison group because an agent picked for coaching after a bad month will usually have a better one anyway. We explain why weak agents’ scores rise without help and how to measure what coaching added.
Forty reviewed calls per agent a month is about ten times the usual sample. Few QA budgets stretch that far, so most teams ask five calls to answer all five questions.
What changes when every call is scored
Conversational Intelligence scores every call on your own scorecard, so an agent’s month rests on all of their calls instead of five, and each score links to the sentence, timestamp and speaker behind it. Coaches still pick two or three calls to listen to, now from the whole month instead of a handful.
The fourth row gets easier too. Coaching gains are measured on the trained skill across all of an agent’s calls before and after the course, and set against agents who weren’t coached.
How many calls is the lowest score on your last QA sheet built on?
Want to see how far your sample is from the full picture? We’ll score a week of your real conversations on your own scorecard, so you can set each agent’s full score beside the one from their QA sample. Read more about Conversational Intelligence.
